You bring your own model API keys either way, so inference spend is a wash. The real build is the control plane around the models: DLP on every prompt, policy-based authorization for agents, compliance-grade audit, integrations, durable orchestration, and above all the specialist team you would have to hire, retain, and keep on call.
This is not a wrapper around a model API. A Kindo-class platform is a product in its own right, and every piece below is table stakes, not gold plating.
Durable agent orchestration, a model gateway with routing and failover, RAG (vector store, embeddings, indexing), code-execution sandboxing, SSO/SCIM, and metering.
OAuth broker, a custom provider catalog, and a fleet of MCP servers that must track every vendor's API changes, forever.
DLP/PII redaction on model traffic, per-resource policy authorization for agents, and identity-linked agent audit. Absent from every general automation platform surveyed; you build them either way.
Every third-party component is pinned, patched, and rebuilt. CVE and breaking-change churn across the stack never stops, and neither can the team that absorbs it.
ML-infrastructure and platform engineers are bid on by AI labs and AI-native startups. Expect 4–6+ month searches, agency fees of 20–25% of first-year base, and 15–20% annual attrition in exactly the hardest seats.
| Role | FTE | Fully loaded / yr |
|---|---|---|
| Tech lead / engineering manager | 1 | $350k |
| Senior backend / platform engineers | 3 | $285k each |
| AI platform engineers model gateway, routing, evals | 1.5 | $380k each |
| Frontend engineers | 1.5 | $255k each |
| DevOps / SRE | 1 | $285k |
| Security engineer DLP, authorization, sandbox review | 1 | $300k |
| Build-phase team | 9 | ~$2.7M / yr |
Year one is pure spend: nothing ships until month 15–18. And it never drops below ~$2.8M a year after that. Model API token spend is excluded from both sides: you bring your own inference either way.
| Year 1 | Year 2 | Year 3 | Total | |
|---|---|---|---|---|
| Payroll (fully loaded) | $4.0M | $2.2M | $2.2M | $8.4M |
| Recruiting + backfill | $0.5M | $0.15M | $0.15M | $0.8M |
| Cloud platform infrastructure | $0.25M | $0.3M | $0.3M | $0.85M |
| Compliance + tooling | $0.1M | $0.15M | $0.15M | $0.4M |
| Total | ~$4.9M | ~$2.8M | ~$2.8M | ~$10.5M |
This is a people problem, not a cloud-bill problem. Nearly nine of every ten dollars go to hiring and retaining a team whose market clears against you. The lean, everything-goes-right floor is ~$7M; the realistic ceiling is ~$13M.
DIY estimates always price the build. They almost never price what comes after.
A sane 24/7 rotation needs five to six engineers who each understand the whole system. You cannot staff half a rotation: ~$1.7–2M/yr just to keep the lights on, before a single new feature.
Model provider APIs and capabilities change quarterly. Every new model, tool, and integration re-opens your DLP, authorization, and audit surface, and someone has to close it again.
The platform carries its own SOC 2, pentest, and audit burden. Evidence collection, tooling, and auditor fees land on your team, every year.
DIY costs roughly $4.9M and 18 months before the first user logs in, then ~$2.8M every year after, contingent on winning a hiring market you are structurally disadvantaged in. Kindo SaaS lists at $250k per year, bring your own inference, and you could deploy this week.